Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs

Name: Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs
Uploaded: 2026-09-30T20:58:27.972674+00:00
Duration: 859 s
Description: The episode discusses research showing that fine-tuning audio LLMs on benign data can degrade safety alignment, increasing jailbreak success rates up to 87.12%. The degradation depends on whether benign data is positioned near harmful content in semantic, acoustic, or mixed embedding spaces. Defenses proposed include filtering training data and using inference-time system prompts.

The episode discusses research showing that fine-tuning audio LLMs on benign data can degrade safety alignment, increasing jailbreak success rates up to 87.12%. The degradation depends on whether benign data is positioned near harmful content in semantic, acoustic, or mixed embedding spaces. Defenses proposed include filtering training data and using inference-time system prompts.

Open the full episode